The AI labs agree to slow down, OpenAI's new model does new things rather than old things better, and a cheap new kind of AI makes decisions instead of text

The heads of nearly every major AI lab backed a plan to pace development and let outside evaluators inside their companies, which moves AI regulation from a shouting match to a negotiation. Meanwhile GPT-6 Astra arrived with a very different set of strengths, near-frontier models got cheap with a data catch attached, and the most useful ideas of the week were small: cheap AI judgment calls, voice, and agents a whole team shares.

• Anthropic, OpenAI, Google, Microsoft and Musk all backed a plan to pace frontier AI, starting with outside evaluators inside the labs
  My take: The week started with a resigning Anthropic researcher saying the labs are 'gambling with our lives', a post seen about 150 million times, and two dozen US politicians calling for legislation within a day. It ended with Anthropic's CEO publishing a specific three-step plan: independent evaluators embedded in every frontier lab now, shared safety standards across democratic countries next, and a global agreement last. Sam Altman committed OpenAI to the evaluators, Elon Musk said 'Dario is right', and Google DeepMind and Microsoft signed on to the direction. Zuckerberg and Nvidia's Jensen Huang took the other side: each lab should slow itself when needed and skip the collective deal. Pacing here does not mean stopping, it means more time between a model being finished and you being able to buy it, and that gap was already the widest it has ever been. For a business owner I think this is mildly good news. One real reason companies delay AI projects is that the tools change every three months, and a steadier release rhythm makes it easier to commit. What I would do: stop waiting for the next model before starting a project, because the next model may now arrive later than you expect, and assume some form of AI regulation lands within a year or two, so keep a simple record of where AI touches customer data and decisions in your business.
• GPT-6 Astra is the best model yet at operating a computer and building in 3D, not clearly better at everyday work, and demand forced OpenAI to pause new 200 dollar subscriptions
  My take: OpenAI's long-awaited model scored 41 percent on the main test of completing tasks by operating a computer, against 31 percent for Anthropic's Fable 5.1 and 18 percent for OpenAI's previous model. People are using it to build playable games, 3D house walkthroughs from a property listing, and interactive explainers in under an hour. But experienced users report little improvement on normal coding and writing, weaker website design than Claude, and one software team saw its effective spend double before moving part of the team back to the older model. Demand was high enough that OpenAI paused new sign-ups to its 200 dollar Pro plan, and Microsoft now plans to triple its data center capacity by 2032. The take here is that subsidised AI is ending, so budget for prices to rise rather than fall on premium plans. Do not switch your daily work to Astra by default. Test it on two things instead: a stable, repetitive workflow that lives in a browser or a desktop app (CRM updates, lead routing, report pulls), and one thing you never had the budget to make, like an interactive product demo or a proposal where the client can change scope and see the timeline and price move.
• A new kind of AI model answers yes or no questions with a probability instead of writing text, at a tiny fraction of the cost
  My take: A startup founded by one of the people behind ChatGPT released Jev, a model that does not write anything. You ask it a narrow question (is this customer angry, is this a buying enquiry, does this draft promise something we cannot deliver) and it returns a number like 0.9, which your software then uses to route, flag or escalate. The company claims 20 to 200 times the speed and 40 to 400 times lower cost than a chat model, and one early tester ran 777 judgments across 37 documents in under a second for about a quarter of a cent. Those are early claims, so treat them with care, but the category makes sense. A large share of office work is reading something and deciding what happens next, and chat models are a slow, expensive way to do that. The pattern to remember is: a chat model proposes, a judgment model decides, ordinary code acts. If you have someone sorting an inbox, triaging support tickets or scoring leads by hand, write down the five questions they are really answering each time. That list is the specification for this kind of tool, and it is useful even if you never buy this particular product.
• Near-frontier AI got very cheap this week, and the price is often your data
  My take: Meta released a model that independent testers rank close to the best from Anthropic and OpenAI at about 55 cents per task, roughly a quarter of the cost of Anthropic's Opus 5. Google shipped its third fast, cheap model in six weeks, DeepSeek released one at 30 cents per million input tokens, and the coding company Cognition built its own model that is 64 percent cheaper than Fable 5.1 for near-equal results. Two warnings. First, analysts showed that Meta's and Google's models score far better on published tests than on a brand new one, so trust your own tasks over any leaderboard. Second, check where your data goes. The free tier of Meta's model exists because Meta may train on what you type. Anthropic reported that a Chinese lab quietly forwarded almost 300,000 of its own customers' requests to Claude in ten days, exposing sensitive corporate and government information. And OpenAI, in a public dispute with a mathematician over a famous solved problem, confirmed it uses de-identified usage data to improve its models, as it says every lab does. Use cheap models for high-volume routine work, that is what they are for. But before anything confidential goes in, read the training and retention terms, turn off training where you can, and prefer business plans where the contract says your data is not used.
• OpenAI opened its live voice model to developers at five cents a minute and bundled its small business connectors
  My take: The voice system inside ChatGPT, which can listen and talk at the same time, handle interruptions and background noise, and complete tasks while the conversation continues, is now available for any software company to build on at five cents a minute plus normal model costs. Expect a wave of phone and voice products over the next few months. If you run a service business that misses calls because you are on the tools, or a support line with a queue, this is the release that makes an AI receptionist worth pricing, so ask your phone or booking software vendor what they have planned. OpenAI also gathered its connectors for Shopify, QuickBooks, Stripe, HubSpot, Gusto, DocuSign and similar tools into one small business collection, and released a finance version with built-in market data and a data agent that answers questions from your company warehouse. If you already pay for ChatGPT and use any of those apps, spend 30 minutes connecting one and asking it a question you normally answer by exporting a spreadsheet. And if you still type every prompt, try voice for a week. The whole industry is building for it.
• The next step for AI agents is one your whole team shares, not one per person
  My take: Almost every agent in use today works for one person, in a private thread nobody else can see. But surveys put solo work at around 40 percent of the day, with the rest spent on coordination, handoffs, meetings and messages. The leading teams have started closing that gap. Anthropic now runs a single shared Claude per Slack channel that everyone can see, redirect and pick up from where a colleague stopped, and it says 65 percent of its product team's code now comes through that shared agent. Y Combinator has asked founders to build the same thing for sales, support, legal and marketing teams. You do not need to wait for those products. Three steps get you ready: find out what each person on your team already runs (you will be surprised by the range), write down the context everyone keeps re-explaining to their own AI and put it in one shared place, then pick one recurring piece of work that several people touch and where you can quickly check whether the result is right. Run one shared agent on that for a month. Choose work where a mistake is cheap to catch, and name one person who owns the result.